Estimating Event-Free Survival From All Causes

What the Product-Limit Estimate Measures

In a competing-risks setting, a person may experience one of several event types, but any one of those events ends the event-free state. The event-free survival function describes the probability that no event of any cause has occurred by a specified time. It answers an all-cause question; it does not, by itself, say which cause occurred or estimate the probability of a particular cause.

The product-limit estimate combines information across event times while allowing some individuals to be right-censored before an event is observed. The essential calculation is to use the number still at risk just before each event time as the denominator, and to count events from all causes in the numerator. Censoring reduces later risk sets, but does not create an event-time factor in the product.

Definition: Let $T$ be the time from a defined origin to the first event of any cause. The event-free survival function is $S(t)=P(T>t)$, the probability of remaining free of every event cause beyond time $t$. At distinct observed event times $t_1<\cdots<t_K$, let $Y_j$ be the number at risk immediately before $t_j$, and let $d_j$ be the number experiencing an event of any cause at $t_j$. The product-limit estimate is $$ \widehat S(t)=\prod_{j:t_j\le t}\left(1-\frac{d_j}{Y_j}\right). $$ If no event time is at or before $t$, the empty product is 1.

The risk set $Y_j$ includes individuals who have not had an event and have not been censored before $t_j$. Under the usual convention for exact event times, an individual with an event at $t_j$ is in the risk set immediately before that time; that individual is then counted among the $d_j$ events. A person censored earlier is not included in later risk sets. If event and censoring times are recorded as tied but their true ordering is unknown, the data-handling convention must be specified rather than silently assumed.

How to Calculate It

For each event time, first find $Y_j$, then count all events at that time to obtain $d_j$. The factor $1-d_j/Y_j$ is the estimated fraction of that risk set that remains event-free through that event time. Multiplying the factors gives the estimated event-free proportion through $t$. Tied event types at the same time contribute together to $d_j$; the cause labels do not create separate factors for the all-cause estimate.

A short justification follows from the product rule for sequential conditional probabilities: remaining event-free through successive event times requires remaining event-free at each one. The observed fraction of the risk set with an event at time $t_j$ is $d_j/Y_j$, so the corresponding estimated fraction without an event is $1-d_j/Y_j$. Multiplying those successive fractions gives the displayed estimator. Censored individuals contribute to the risk set while observed, and then leave it; they are not counted as events.

Worked Example: No Censoring Before the Event Times

Suppose 10 hypothetical participants are followed from a common origin. There are 2 events at time 2, 1 at time 5, and 2 at time 8; no one is censored before time 8. Calculate event-free survival at times 4, 6, and 8.

At time 2, $Y_1=10$ and $d_1=2$. After those events, 8 people remain at risk. Thus at time 5, $Y_2=8$ and $d_2=1$; 7 remain at risk for time 8. At time 8, $Y_3=7$ and $d_3=2$.

At time 4, only the factor for time 2 is included: $ \widehat S(4)=1-2/10=8/10=0.8000 $. At time 6, include the factors for times 2 and 5: $ \widehat S(6)=(8/10)(7/8)=7/10=0.7000 $. At time 8, include the factor for time 8 as well: $ \widehat S(8)=(8/10)(7/8)(5/7)=5/10=0.5000 $.

The estimate at time 8 is 0.5000, or 50.00%. In this hypothetical data set, half of the original group remains event-free through time 8 according to the product-limit calculation.

Worked Example: Censoring Changes Later Risk Sets

In a second hypothetical group of 12 participants, 2 events occur at time 1, 2 participants are censored at time 2, 1 event occurs at time 3, 1 participant is censored at time 4, and 2 events occur at time 6. Assume the exact event and censoring times are known and there are no other events before time 6. Calculate the estimate at times 5 and 6.

At time 1, $Y_1=12$ and $d_1=2$, giving a factor $10/12$. Two people are then censored at time 2, so they are removed from the subsequent risk sets without contributing an event. At time 3, $Y_2=8$, $d_2=1$, and the factor is $7/8$. The censoring at time 4 leaves 6 participants at risk at time 6. There $Y_3=6$, $d_3=2$, and the factor is $4/6$.

At time 5, the time-6 factor is not included: $ \widehat S(5)=(10/12)(7/8)=70/96=0.7292 $. At time 6: $ \widehat S(6)=(10/12)(7/8)(4/6)=280/576=0.4861 $.

The estimate is 0.4861 at time 6. The two censorings do not count as events, but they matter: without removing them, the later risk-set sizes and event fractions would be calculated incorrectly.

Multiple Causes and Simultaneous Events

For this estimate, the event definition is “an event of any cause.” If $d_{jk}$ is the number of events from cause $k$ at time $t_j$, then the all-cause count is $d_j=\sum_k d_{jk}$. This sum is used once in the factor for $t_j$. The calculation does not split the all-cause event-free estimate into cause-specific probabilities.

Worked Example: Combining Event Types at a Shared Time

A hypothetical cohort begins with 10 participants. At time 2, 1 event of cause A and 1 of cause B occur. At time 5, 2 events of cause A and 1 of cause B occur. At time 7, 1 event of cause B occurs. Assume there is no censoring before time 7. Find the product-limit estimate at time 7.

At time 2, the total event count is $d_1=1+1=2$, with $Y_1=10$, so the factor is $1-2/10=8/10$. There are 8 at risk at time 5. The all-cause count there is $d_2=2+1=3$, giving $1-3/8=5/8$. Five remain at risk at time 7; there $d_3=1$, so the factor is $1-1/5=4/5$.

Therefore, $ \widehat S(7)=(8/10)(5/8)(4/5)=4/10=0.4000 $. The estimate is 40.00% event-free through time 7. The cause labels were needed to verify the total event counts, but the all-cause calculation uses their sum at each time.

Assumptions and Checks

The arithmetic is simple, but interpreting the estimate as event-free survival in a target population requires suitable data and assumptions. In particular, the event times and censoring indicators must be recorded accurately, and every event that ends the defined event-free state must be counted, regardless of its cause. Censoring must be independent of the event time, at least within any strata or covariate groups used for the estimate. Informally, among individuals still event-free and observed, those censored at a given time should not have a different future event-time distribution from those who remain under observation.

The simple formula also presumes a common time origin and that individuals enter observation at that origin. The risk set must be constructed for the actual observation scheme; the formula as written does not adjust for delayed entry. If censoring is informative, or if important groups have different event-time distributions but are pooled, the resulting estimate may not represent the intended population survival function. In addition, a cause must not be treated as censoring merely because it is not the cause currently of interest: for this all-cause quantity, every event ends event-free time.

Several checks catch common calculation errors:

At every event time, confirm $0\le d_j\le Y_j$, and check that the risk set includes only people still observed and event-free immediately beforehand.
Between event times, censoring changes later risk sets but does not directly change $\widehat S(t)$. The estimate changes only at event times.
Each factor lies between 0 and 1, so the estimate cannot increase over time. A rising estimate indicates a coding or calculation error.
Report the number at risk near the time being interpreted. A flat estimate after the last observed event does not mean follow-up continued or that the true event-free probability is known to remain constant.

The independence of censoring is not established merely because the event-free estimate looks smooth. Compare censoring patterns across relevant groups, investigate whether follow-up ends for reasons related to prognosis, and explain the censoring mechanism in the analysis. Such checks can reveal concerns, but cannot prove independent censoring.

What This Estimate Does—and Does Not—Compare

A raw proportion of the original group with no recorded event by time $t$ is suitable only when event status through $t$ is known for everyone, or when missing follow-up is handled under additional justified assumptions. Counting people lost to follow-up as event-free, or dropping them without considering why they were censored, can distort that proportion. The product-limit estimate instead uses the observed follow-up by updating risk sets at each event time.

A parametric survival model is another approach: it specifies a distribution or functional form for event times and may provide smoother estimates or allow covariate modeling, but its conclusions depend on those modeling choices. The product-limit estimate is nonparametric in the event-time distribution and is useful as a direct summary, subject to its censoring and observation assumptions. Neither approach turns this all-cause event-free quantity into the probability of a particular event cause. Estimating cause-specific event probabilities requires a different calculation.

Common pitfall: A researcher sees two event types and calculates separate product factors for them, or treats one type as censoring while estimating event-free survival. That does not answer the all-cause question. For remaining free of every event cause, add the event counts across causes at each event time, use the full risk set immediately before that time, and include one factor $1-d_j/Y_j$.

Check Your Understanding

Answer the questions, then check the solutions below.

At an event time, $Y_j=20$, with 2 events of one cause and 3 of another. What event count and product factor are used for all-cause event-free survival?
Why does a participant censored between two event times affect the later calculation but not create a product factor at the censoring time?
Suppose $Y_1=10,d_1=2$, followed by $Y_2=8,d_2=2$. Calculate the estimate just after the second event time.
Name one assumption needed to interpret the product-limit estimate as population event-free survival, and give a practical concern that could violate it.

Solutions. (1) The total is $d_j=2+3=5$, so the factor is $1-5/20=0.7500$. (2) The participant contributes to risk sets while observed, then is removed from later risk sets; censoring is not an event, so it contributes no event-time factor. (3) $\widehat S=(1-2/10)(1-2/8)=(8/10)(6/8)=0.6000$. (4) Independent censoring is one assumption; for example, participants who leave follow-up because their health is worsening may have different future event times from those who remain observed.